Question I have a question about using a poca trainer.

I studied and understood the ppo algorithm, the basic algorithm of ml-agents. However, I am not sure how the poca trainer was implemented in ML-Agents. I have read two papers on multi-agent. Is it right that poca trainer centralize policies and take action separately? I need details. Thank you!

 

Hi @xogur6889,
The POCA algorithm was developed by the ML-Agents research team. They're working on an arxiv submission, but it's not ready yet. I'll see if there are any other details that they can share sooner...

 

I talked to the research folks some more. The high-level explanation is that POCA trains a group of agents to maximize a shared common reward. It also supports the removal/addition of agents during runtime, which I believe is not handled by other multi-agent trainers.

The POCA trainer trains policies in a centralized way, but each policy acts independently during inference.

 

Thank you for your help. I look forward to other details.

 

I didn't know about this fact. Thank you, but I wonder how you implemented it. Is it possible?

 

HI @celion_unity,
Is there anything new?

 

The team is still working on a submission to arxiv, and there's nothing else to share right now. When the arvix paper is available, I'll update you.

 

Thanks!!! I'll just wait. hahaha

 

Hi! Is there any news already?

 

hello，about this paper，there is RSA information. But I cant find the setting in ....yaml